Gemini CLI v0.38 Ships Subagents and Context Compression

Gemini CLI v0.38 shipped in mid-April with two features that change how the tool handles long, complex sessions: a subagent system for delegating specialized work, and a context compression service for keeping long conversations focused.

The last time we covered Gemini CLI was at v0.37, which introduced session chapters and dynamic sandbox expansion. v0.38 builds on that with delegation and memory management, the two capabilities that matter most as tasks grow in scope.

Subagents: Delegating Work to Specialized Agents

The headline feature in v0.38.1 is subagents. When you’re working through a complex task in Gemini CLI, your main agent can now spawn specialized subagents to handle specific subtasks without cluttering the primary context.

Each subagent runs in a separate context loop with its own system prompt, its own persona, and optionally a restricted or specialized tool set. When the subagent finishes its work, it returns only the synthesized result to the parent session. The intermediate work, all the context the subagent accumulated while investigating a codebase or analyzing a document, doesn’t land in your main conversation history.

This is a meaningful architectural shift. Until now, Gemini CLI handled all work in a single session context. If you asked the agent to read through a large codebase to understand how authentication worked before implementing a new feature, all of that investigation would pile into the same context as your implementation work. With subagents, you can hand that investigation off to a dedicated agent that reads, analyzes, and returns a summary.

Gemini CLI ships with three built-in subagents. The generalist has access to all tools and handles open-ended tasks. The cli_help agent specializes in questions about Gemini CLI itself. The codebase_investigator is designed for exploring codebases, mapping architecture, and understanding how systems fit together. That last one is the most useful for active development work: it reads widely, thinks carefully, and returns structured answers rather than raw output.

Google announced the feature on the Developers Blog with an emphasis on how subagents let your main agent stay focused on the high-level task while specialized experts handle the detail work. The GitHub discussion thread at #25562 has additional technical detail on how the delegation protocol works.

Context Compression

The second major addition is a dedicated ContextCompressionService, implemented in v0.38.0. It runs as a background service that monitors your conversation history and distills it as the session grows.

Long CLI sessions have had a consistent problem: context bloat. After an hour of work, the early parts of the conversation become noise rather than signal. The model spends tokens on context that’s no longer relevant to the current task. Context compression addresses this by actively pruning and summarizing earlier conversation history, keeping the active context focused on what’s actually relevant to what you’re doing now.

This works alongside the Chapters feature from v0.37, which organizes sessions into labeled topic groups. Compression can now distill earlier chapters into summaries rather than retaining every tool call in full fidelity.

Other Changes in v0.38

The release also includes background process monitoring tools, giving the agent better visibility into asynchronous tasks it has started. UI stability improvements landed as well, with the “Terminal Buffer” mode from v0.37 refined further. Browser agent support improved, and Windows reliability got several fixes.

The upcoming v0.39.0-preview.0 refactors the subagent tools further, consolidating them into a single invoke_subagent tool and removing the legacy individual wrappers. If you’re using subagents in v0.38, the interface will change in the next stable release.

Updating

Gemini CLI v0.38.2 is the current stable release as of April 17. Update with npm install -g @google/gemini-cli or npx @google/gemini-cli update depending on how you installed it. The full changelog is at geminicli.com/docs/changelogs.

Sources: Google Developers Blog, Gemini CLI Changelog, InfoQ, GitHub Discussion #25562